Classical Image Processing

Classical image processing uses fixed operations such as convolution, thresholding, morphology, and geometric transforms. It is not obsolete: it is often the most auditable part of an OCR pipeline, a preprocessing step before feature extraction, or a sanity baseline for semantic segmentation.

Convolution and gradient filters

For a grayscale image $I$ and kernel $K$, 2D convolution computes

$$(I\ast K)[u,v]=\sum _i\sum _{j}I[u-i,v-j]K[i,j].$$

The Sobel operator estimates horizontal and vertical derivatives with small kernels, for example

$$K_x=\left[\begin{array}{ccc}-1& 0& 1\\ -2& 0& 2\\ -1& 0& 1\end{array}\right].$$

Edges become large responses because neighboring intensities differ strongly; flat regions cancel out.

Worked example

This snippet applies a Sobel-style horizontal edge filter to a toy image and reports the gradient response and maximum edge magnitude.

import numpy as np
from scipy.signal import convolve2d
 
img = np.zeros((5, 5)); img[:, 3:] = 10
sobel_x = np.array([[-1,0,1],[-2,0,2],[-1,0,1]])
gx = convolve2d(img, sobel_x, mode="same", boundary="symm")
print("image")
print(img.astype(int))
print("sobel_x")
print(gx.astype(int))
print("max_abs_gradient", int(np.abs(gx).max()))

Observed output:

image
[[ 0  0  0 10 10]
 [ 0  0  0 10 10]
 [ 0  0  0 10 10]
 [ 0  0  0 10 10]
 [ 0  0  0 10 10]]
sobel_x
[[  0   0 -40 -40   0]
 [  0   0 -40 -40   0]
 [  0   0 -40 -40   0]
 [  0   0 -40 -40   0]
 [  0   0 -40 -40   0]]
max_abs_gradient 40

Only the vertical intensity jump produces a strong derivative. The same array contract depends on correct image representation: channel order, dtype, and padding convention all change the result.

Caveats

Fixed thresholds break under lighting changes, blur, and sensor shifts. Morphology can remove the small structures that medical image analysis cares about. Classical steps should be versioned and benchmarked like learned models, not treated as harmless preprocessing.

References

• SciPy documentation: scipy.signal.convolve2d
• SciPy documentation: scipy.ndimage.sobel

Nav

Section — Computer Vision

← Image Representation Feature Extraction →